Accelerating Game Asset Creation with Fine-Tuned Diffusion Models

rovio · Media & Entertainment · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Game art production

Rovio needed to help artists produce high-volume Angry Birds assets and variations without losing brand control.

Constraints:

• Brand consistency
• Artist control
• Multiple user workflows from simple to advanced

Architecture

Rovio built Beacon Picasso as a generative-art platform for game teams rather than a single image model. Artists help curate and caption small high-quality style datasets; SageMaker runs many fine-tuning jobs against base diffusion models and writes safe-tensor artifacts to S3; artists review training progress to pick usable checkpoints; workflow tools and multiple UIs expose the models; EC2 G6e instances serve fast image generation inside Rovio's VPC; and artists remain at the center to prompt, select, clean, upscale, layer, and integrate AI-generated art into actual game assets.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  beacon-picasso[Beacon Picasso] --> artist-facing-tools[Artist-facing tools]
  artist-facing-tools --> slack-bot[Slack Bot]
  artist-facing-tools --> workflow-tools[Workflow tools]
  artist-curated-dataset[Artist-curated dataset] --> sagemaker-training-jobs[SageMaker training jobs]
  base-diffusion-models[Base diffusion models] --> sagemaker-training-jobs
  sagemaker-training-jobs --> s3-model-artifacts[S3 model artifacts]
  s3-model-artifacts --> artist-checkpoint-review[Artist checkpoint review]
  s3-model-artifacts --> fine-tuned-diffusion-models[Fine-tuned diffusion models]
  fine-tuned-diffusion-models --> workflow-tools
  workflow-tools --> ec2-g6e-inference[EC2 G6e inference]
  workflow-tools --> asset-finishing-pipeline[Asset finishing pipeline]
  beacon-picasso --> fine-tuned-diffusion-models

Components

Data flow

The team picks a base diffusion model, builds an artist-curated dataset, fine-tunes models in SageMaker, stores checkpoint artifacts in S3, has artists choose the best model, loads the artifact into tools/workflows, runs inference on EC2 G6e instances, and lets artists turn generated images into game assets through cleanup, expansion, upscaling, layering, and UI integration.

Why this architecture

Rovio chose a platform-plus-workflow architecture because generic image tools could not meet Angry Birds brand constraints, and artists had different needs for control. Fine-tuning on proprietary style data gave brand fit, SageMaker/EC2 provided training and fast inference, and multiple tools kept artists in creative control instead of forcing one interface.

Models used: fine-tuned diffusion models

Scale

• SageMaker let Rovio run thousands of training jobs in parallel and get results in hours.
  “So we can run thousands of trainings in parallel and get the results in a couple of hours. Why we need those of trainings, as you can imagine, if you have worked with machine learning, the first training is not always the best one.”transcript, stated
• EC2 G6e instances made image generation happen in seconds for artists.
  “And all this is powered by EC2 G6e instances, which makes this generation happen in seconds. So our artists don't get frustrated because they don't have to wait, but okay, we have images in our style but we need to turn them into game assets.”transcript, stated
• The season-pass workflow achieved more than 80% production-time savings.
  “I think at the beginning I mentioned that they were able to achieve more than 80% of production time savings and that is for the season pass background use case where they initially used to take 20 days and now they could achieve the same outputs in four days.”transcript, stated

Alternatives considered (roads not taken)

Generate flagship Angry Birds characters first

Rejected because: Flagship characters had strict brand proportions and expression rules that early models failed to satisfy.

Use generic base diffusion models as-is

Rejected because: Base models produced good realistic images, but not Rovio's style.

One simple interface for every artist

Rejected because: Artists varied from wanting simple prompt interfaces to wanting complex workflow controls.

Let ML engineers curate and caption the style data

Rejected because: Artists had the vocabulary and trust relationship needed to describe the visual style well.

Patterns

fine-tuning-adaptation (model-operations)

Rovio trained custom image models on Angry Birds art rather than relying only on base diffusion models.

human-in-the-loop-gate (creative-workflow)

The tools are designed for artists to use and control rather than fully replacing creative judgment.

Tradeoffs

Apply generative art first to flexible asset classes.

Chose Secondary characters, backgrounds, and locations where exact control was less critical over Primary brand characters that required exact proportions and expression.

Rationale: The system produced useful creative acceleration when the art direction tolerated variation.

Put artists in charge of model selection.

Chose Human artist review of training-progress videos and candidate model artifacts over Selecting checkpoints solely from automated training outputs.

Rationale: The quality bar was visual and brand-specific, so the people who owned the art direction had to choose the usable model.

Where to apply generative art first.

Chose backgrounds, secondary characters, and non-brand-essential assets over primary Angry Birds characters with strict brand rules.

Rationale: Rovio got value where style variation was acceptable and avoided use cases requiring exact character control.

How to support artists with different AI comfort levels.

Chose multiple tools and workflow surfaces over one universal interface.

Rationale: Some artists wanted simplicity while others wanted complex control, so the platform exposed different entry points.

Gotchas (what bit them)

Early character fine-tunes looked promising to engineers but failed the studio's art-quality bar.

Root cause: Generating a flagship character required brand-specific proportions, expression, and style control that generic outputs did not capture.

Resolution: Rovio kept artists in the evaluation loop and shifted the AI use cases toward assets where creative variance was acceptable.

Artists initially struggled to control diffusion outputs for exact branded assets.

Root cause: The models were stochastic and difficult to steer for precise Angry Birds character requirements.

Resolution: The team focused on use cases with more stylistic flexibility and taught workflows such as seed locking and prompt iteration.

Evaluation

Rovio measured production-time savings for specific art workflows.

Metrics: production time saved

Outcomes

• Specific asset workflows saw roughly 80% production-time savings.

The tools supported multiple interfaces for different artist needs.